Research map / Vision-language and multimodal / Robot action and navigation
SCOUT: Planning under Occlusion via Object-Centric World Model Rollouts
Amir Hossain Raj, Xuesu Xiao
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SCOUT, an occlusion-aware robot planning framework that combines high-level deliberation with object-centric world model rollouts to plan manipulation under partial views in cluttered scenes.
Problem
Robots manipulating in cluttered spaces must plan from partial views where foreground objects occlude the target and free space; reactive policies miss hidden constraints while exhaustive search is computationally prohibitive in dense clutter.
Approach
Bridges high-level deliberation with grounded physical rollouts from an object-centric world model to reason about occluded constraints without exhaustive search.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use multi-view instead of single-view input (ablation) | SCOUT · simulated shelf | avg success (%): 69.7 -> 80.8 | Improved |
| add relational projection of planning state (ablation) | SCOUT · simulated shelf | avg success (%): 68.1 -> 80.8 | Improved |
| add MCTS rollout search instead of greedy depth-1 (ablation) | SCOUT · simulated shelf | avg success (%): 70.6 -> 80.8 | Improved |
| add LLM proposals to geometric proposals (ablation) | SCOUT · simulated shelf | avg success (%): 74.6 -> 80.8 | Improved |
| increase MCTS search depth from 2 to 3 or 4 (ablation) | SCOUT · simulated shelf | avg success (%): 80.8 -> 81.5 | No clear effect |
| increase search depth from 1 to 2 (ablation) | SCOUT · simulated shelf | avg success (%): 70.6 -> 80.8 | Improved |
| replace GHOST initial state with oracle geometry (ablation) | SCOUT · simulated shelf | avg success (%): 80.8 -> 87.8 | Improved |
| re-observe after every primitive | SCOUT · simulated shelf | hidden-scene failures (%): 7.6 -> 3.3 | Improved |
| re-observe after every primitive | SCOUT · simulated shelf | episode time (s): 10.4 -> 24.2 | Hurt |
| multi-view instead of single-view GHOST reconstruction (ablation) | GHOST · held-out reconstruction | hidden-object recall (%): 58.9 -> 72.6 | Improved |
Closest papers by meaning
- SyncWorld: Visual Calibration Enables World Models as Zero-Shot Simulators (NeurIPS 2026 · similarity 0.86)
- FLIP: Flow-Centric Generative Planning as General-Purpose Manipulation World Model (ICLR 2025 · similarity 0.85)
- PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking (NeurIPS 2026 · similarity 0.85)
- MolmoMotion: Forecasting Point Trajectories in 3D with Language Instruction (NeurIPS 2026 · similarity 0.85)
- Sparse Planning in Visual World Models via Cost Gradients (NeurIPS 2026 · similarity 0.85)
- Efficient Robotic Policy Learning via Latent Space Backward Planning (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
